Maximum-likelihood affine cepstral filtering (MLACF) technique for speaker normalization

We present a novel technique of minimizing the acoustic variability of speakers by transforming the features extracted from the speaker’s data to better fit the recognition model. The concept of maximum-likelihood affine cepstral filtering (MLACF) will be introduced for feature transformation, along with solutions for the transformation parameters that maximize the likelihood of the test data with respect to a given recognition model. It is shown that for log-concave distributions, the solution of the MLACF problem can be obtained using convex programming. HMM-based digit recognition on the TIDIGITS database is presented to demonstrate the flexibility of the transformation in compensating for large acoustic mismatches between the speakers in the training and test database. In addition, it will be shown that the technique requires estimation of far fewer transformation parameters compared to existing techniques, thus allowing fast, real-time compensation.